BUILD AN EFFICIENT DEEP LEARNING MODEL TO RECOGNIZE SKIN DISEASE BASED ON SELF-KNOWLEDGE DISTILLATION

Abstract

Skin cancer is currently one of the most common diseases with an increasing incidence. Therefore, early prediction or recognition of skin diseases is currently of great interest to researchers around the world, especially in the ISIC skin disease classification contests of 2017, 2018, 2019 and 2020. In this paper, we propose an effective new approach to solve the problem of skin disease identification based on self-knowledge distillation. Our method exploits and minimizes the difference between two probability distributions from two different versions of the same input image. The experiment results performed with the ResNet-50 network have shown that our proposed approach outperforms the state of the art proposed methods on standard datasets such as HAM10000, ISIC 2017 and ISIC 2019. Specifically, our method achieves 0.987 in terms of AUC on the HAM10000 dataset and 0.960 in terms of AUC, 0.901 in terms of accuracy, 0.910 in terms of sensitivity, and 0.866 in terms of specificity on the ISIC 2017 dataset.